Pika H3 Flex 6-Step Sampling Explained: How Pika Cut MiniMax H3 from 50 Steps to 6
PikaH3Flex EditorialHow Pika H3 Flex generates video and audio in six steps: PDMD distillation, ReFL and DPO reward LoRAs, the sampler settings you must keep in ComfyUI, and what the speed-up really means.
Pika H3 Flex makes a full clip — picture and synchronized audio — in six denoising steps, where the MiniMax H3 model it is based on uses 50. Pika got there with a distillation method called Projected Distribution Matching Distillation (PDMD), then added two reward-tuned LoRAs that were merged into the final adapter. This article explains each stage from Pika’s technical report, what it means for speed and quality, and the settings that keep the six-step schedule working when you run the open weights.
Why steps matter
Diffusion video models start from noise and remove it over a sequence of steps. Each step is a full pass of a very large transformer — 33B parameters for MiniMax H3 — over every frame and the audio track. Cutting steps from 50 to six removes most of that compute, which is why Flex can make several takes in the time one 50-step render takes.
Stage 1: a 50-step teacher
Pika started from the official, guidance-distilled MiniMax H3 and fine-tuned it on an in-house audio-video dataset, keeping the 50-step schedule. That fine-tuned model is the teacher: it defines the quality the fast model should match.
Stage 2: distill to six steps with PDMD
The student learns to produce the teacher’s output distribution in six steps.
- Pika began with Distribution Matching Distillation (DMD) (Yin et al., 2024), which trains the student so its samples match the teacher’s distribution rather than copying it step by step.
- Standard DMD was unstable in Pika’s experiments, so it switched to Projected DMD (PDMD) (Wang et al., 2026), which Pika says substantially improved stability and output quality.
- The student generates joint audio and video in six steps — the soundtrack is distilled together with the picture, not added later.
Stage 3: reward LoRAs, merged
Starting from the six-step model, Pika trained two LoRA adapters:
- ReFL (Reward Feedback Learning) — pushes aesthetic quality and visual clarity using differentiable reward feedback.
- DPO (Direct Preference Optimization) — learns from preference pairs about instruction following, realism and visual quality, with human pairwise judgments used to calibrate the automated ones.
The two were merged into one adapter, so they add no steps and no inference cost. That merged adapter is what you download as Pika H3 Flex — see the Pika H3 Flex LoRA guide.
What Pika did not publish
The report does not disclose the sigma values, timestep schedule, guidance scale, loss weights, learning rates, dataset size or training time. The working sigma list ships in ComfyUI’s templates and the repo’s Diffusers examples, so copy it from there instead of guessing.
The speed numbers, read carefully
| Claim | What it measures |
|---|---|
| 15 s 768p clip with audio in “as little as ~7 seconds” | Transformer compute in a controlled benchmark |
| 3.409 s → 2.129 s per 5 s 768p clip (1.60× faster) | Full inference on eight B300 GPUs after kernel work |
| ~1.8× faster than Fal H3 Max on average | Latency across text/image/reference-to-video |
None of these is the time you wait in an app, which adds queueing, VAE decoding, upscaling and file transfer. On a single local H100 the official CLI is far slower than Pika’s eight-GPU serving stack. Comparisons with H3 Max are in Pika H3 Flex vs H3 Max.
The upsampler is a separate model
The 1080p, 2K and 4K outputs do not come from more sampling steps. Pika runs a separate 1B-parameter pixel-space transformer in a single pass for 1×–4× upscaling; the first 1080p frame is ready after about 0.17 s on one GPU. It is Pika-hosted only — the open weights stop at 768p.
Keeping six steps working in ComfyUI and Diffusers
The distilled schedule is fragile. From the ComfyUI tutorial and the repo:
- Six steps with the template’s
ManualSigmaslist (at least seven values). - Sampler
lcm, CFG 1.0 viaBasicGuider— no extra classifier-free guidance. - LoRA strength 1.0.
- Do not switch to generic six-step Euler; Pika says it is not equivalent.
- Do not stack turbo or Lightning LoRAs — the model is already distilled.
Raising the step count does not make it “better” the way it does for a normal model; the student was trained for this exact schedule.
Does six steps cost quality?
On Pika’s automatic metrics, Flex scores at or above both the 50-step MiniMax H3 and Fal H3 Max on visual and audio quality, with alignment roughly tied. Human raters still preferred other systems for physics, lighting and action-heavy shots. In short: the distillation plus reward tuning recovered most of the quality, and the speed makes it practical to iterate until a take works.
You can see six-step generation in action without any setup: write a prompt in Pika H3 Flex text to video.
FAQ
What does “6 step” mean in Pika H3 Flex? The model removes noise in six sampling steps instead of MiniMax H3’s 50, producing video and audio together.
How did Pika reduce the steps? By distilling a fine-tuned 50-step teacher into a six-step student with Projected Distribution Matching Distillation, then merging ReFL and DPO reward LoRAs.
Can I use more steps for higher quality? Not reliably. The adapter is trained for its six-step schedule; changing steps, sampler or CFG moves you off it.
Is the 7-second figure real? It is Pika’s controlled benchmark of transformer compute for a 15-second 768p clip, not end-to-end wait time.